Excerpts from Gershom B's message of 2016-07-13 18:01:50 -0700:

On July 13, 2016 at 7:32:47 PM, Edward Z. Yang (ezyang@mit.edu) wrote:

The general notion sounds good to me. I’m semi-indifferent between (1) and (2) though conservatively lean towards the latter.

...

- The Travis build must always be green. We should prioritize adding more tests for things we care about. Look into regular Hackage tests. - PR all your changes (so that you can check that Travis is green), try to get approval for big changes but BE BOLD.

I’d add

- I don’t break backwards compat for APIs without notice and discussion, and don’t break backwards compat for any element of cabal files without lots of discussion.

(and I don’t know what tests would be necessary to help notice this, but they’d probably be useful!)

Simple, we need Hackage-level CI. Here are few possible things to test: 1. Can we parse all of Hackage? 2. Can we build all of Stackage? https://groups.google.com/forum/#!msg/haskell-stack/oi-VJrAIJbE/FjAPVZTUAQAJ Attempting to build all of Hackage is dicey business because it is a combinatorial explosion, and you do not know what is supposed to build. But Stackage is supposed to build. Test that 3. Can we build all Setup.hs scripts on Hackage? This lets us know which APIs in Cabal matter, and which ones we can change. We'll need to establish base truth for this. It is a travesty that this infrastructure does not exist; it may even be possible to do (1) and (3) on Travis, as they should not take too long to do. Edward

On 2016-07-14 at 03:13:31 +0200, Edward Z. Yang wrote: [...]

...

- I don’t break backwards compat for APIs without notice and discussion, and don’t break backwards compat for any element of cabal files without lots of discussion.

[...]

3. Can we build all Setup.hs scripts on Hackage? This lets us know which APIs in Cabal matter, and which ones we can change. We'll need to establish base truth for this.

While we shouldn't break the API without a good reason, we added custom-setup/setup-depends for the very reason to finally be able to break Setup.hs. Now we can specify `Setup.hs`'s dependencies with version constraints accurately, just like we already do with other library dependencies via build-depends. Moreover, cabal-install has a generous implicit upper bound on the Cabal lib version: -- The idea here is that at some point we will make significant -- breaking changes to the Cabal API that Setup.hs scripts use. -- So for old custom Setup scripts that do not specify explicit -- constraints, we constrain them to use a compatible Cabal version. -- The exact version where we'll make this API break has not yet been -- decided, so for the meantime we guess at 2.x. cabalCompatMaxVer = Version [2] [] So criteria 3 should be seen in the light of custom-setup/setup-depends. Otoh, maybe we should collect and batch up all ideas (including refactoring/aesthetic changes) to break the Cabal API for a big Cabal 2.0 rupture... :-)

It is a travesty that this infrastructure does not exist; it may even be possible to do (1) and (3) on Travis, as they should not take too long to do.

So far, the criteria mentioned address mostly regressions. What about changes extending the .cabal syntax? There's frequently proposals to extend the features of .cabal files. I happen to propose extensions myself from time to time (motivated by the issues I see as Hackage Trustee). Since changes to .cabal are effectively changes to the CABAL specification future generations will have to cope with, I'd propose that changes affecting or introducing new .cabal syntax/semantics should go through a proposal process. I.e. write up a specification/proposal outlining motivation (i.e. what problem does this solve), specify what the changes are exactly (syntax & semantics), what the consequences are, and so on. Then we inevitably need some criteria to decide whether a proposal is accepted and approved for implementation. This could be formally in the hands of the core library committee together with the Hackage trustees (since we do have quite some experience with .cabal files and care a lot about the Hackage ecosystem). -- hvr

On 14 Jul 2016, at 09:54, Edward Z. Yang wrote:

Excerpts from Herbert Valerio Riedel's message of 2016-07-13 23:40:06 -0700:

...

I.e. write up a specification/proposal outlining motivation (i.e. what problem does this solve), specify what the changes are exactly (syntax & semantics), what the consequences are, and so on.

Then we inevitably need some criteria to decide whether a proposal is accepted and approved for implementation. This could be formally in the hands of the core library committee together with the Hackage trustees (since we do have quite some experience with .cabal files and care a lot about the Hackage ecosystem).

Why can't we release experimental changes to the Cabal specification? Neither setup-depends nor convenience libraries went through any formal proposal mechanism. If a feature is good people will start using it and we will have to continue supporting it. If a feature is bad/not useful we can yank it from the next release; anyone using an experimental feature like this isn't trying to maximize their Cabal compatibility anyway. And any change to the format with cabal-install can't parse is not going to be supportable via custom-setup anyway.

You gave a reason yourself: - Can we parse all of Hackage? With my industrial hat on, I don’t like the idea that I might need older Cabal, to be able to install some package. OTOH new-build could install older Cabal. So we’d need to maintain older Cabals, to be build able with newer GHCs, that’s not fun, but not impossible. I agree, that we could have discussed setup-depends more, hopefully no-one disagrees about them (and dcoutts took very conservative approach there). Also taking them out of Cabal, doesn’t mean that package becomes unbuildable, there will be just more luck involved. About convenience libraries I agree even more. We should discussed them more. There are questions I’d might to ask, but I guess it too late Or maybe not too late, but for another thread: what is convenience library name clashes with the library on package, say I define convenience library fancylib-lens, which is on Hackage, but maybe also after I made my package? I guess, some shadowing happens? Is there a warning? But we have quite formal feeling proposals about `provides`, and `or-dependencies / multi-way-flag / depends`, and formal process would help to get them thru (or say no, with an explanation)

So how about this script, which is how web standards work:

Cabal changes need to be informally proposed and discussed. But the advancer can also write a PR and get it merged in as an experimental extension, to be trialed for the next major release series. Eventually it gets standardized.

Edward

About the original topic, I’d require “ a history of quality contributions”, so something between (2) or (3). I don’t like many rules, also making fair but non-abusable ones is difficult, but as we have this conversation, the outcome should be well defined (also how we change the rules). As with (2) does every accepted PR counts, like single character typo fix in the documentation? And I don’t know how to define “quality contribution” :( (Maybe I’m just to scared about dramas happened elsewhere) Related to this, a thought I had for some time: I’d like to split Cabal and cabal-install repositories. Pros: - `cabal-install` can be much more liberal in giving out commit bits, as you cannot make too much damage here. - we already have `hackage-security` as a dependency between them. - the CI setup could be simplified, resulting into - their tests could do more things: `Cabal` parses all `Hackage`, changes to `cabal-install` doesn’t need to do it, and vice-versa, fixing ui of cabal-install doesn’t need to do package tests of Cabal. - their other processes could be different too. Cons: Simultaneous development is almost impossible. And this is almost a show-stopper. - `cabal-install` would need to lag one release behind `Cabal` (development agains released `Cabal`). This could work as `Setup.hs` should work to install packages requiring newer Cabal (isn’t it how it works nowadays?). There will be a problem when newer `Cabal` won’t accept build-plan created by solver in older `cabal-install` (proposals mentioned above), but I hope it won’t be a huge issue. - To make this work we’d need to release more Cabal major versions, so there aren’t many breaking and intrusive changes accumulating. But that’s good thing, isn’t it? Or it could be a quality win to separate them. After all `cabal-install` is not the only consumer of `Cabal`. `stack` uses it very limitedly atm though (i.e. to extract dependencies and maybe verify install plan?) I don’t know what `ghc-pkg` does. IMHO we should develop Cabal-the-library with highest standards (starting with having detailed CHANGELOG) for library development. Now we might have been too focused on developing Cabal for cabal-install. Cabal should provide quality plumbing support for “everyone”. (if we could start with a clean desk, i’d make Setup.hs and cabal-install commands different, that’s one confusing part). The separation into different repositories would make things a bit clearer: Cabal-the-library and cabal-install are different things. (For example: cabal-install doesn’t parse *.cabal files). - Oleg

Edward Z. Yang writes:

Excerpts from Herbert Valerio Riedel's message of 2016-07-13 23:40:06 -0700:

...

I.e. write up a specification/proposal outlining motivation (i.e. what problem does this solve), specify what the changes are exactly (syntax & semantics), what the consequences are, and so on.

Then we inevitably need some criteria to decide whether a proposal is accepted and approved for implementation. This could be formally in the hands of the core library committee together with the Hackage trustees (since we do have quite some experience with .cabal files and care a lot about the Hackage ecosystem).

Why can't we release experimental changes to the Cabal specification? Neither setup-depends nor convenience libraries went through any formal proposal mechanism. If a feature is good people will start using it and we will have to continue supporting it. If a feature is bad/not useful we can yank it from the next release; anyone using an experimental feature like this isn't trying to maximize their Cabal compatibility anyway.

Important nitpick: to do that, experimental features should be marked as experimental and this policy should be made explicit. That'd be kinder to the one user who starts using it without realizing. I guess that goes without saying (probably in ways I don't realize), but hey :-D